Source: OpenAI Deep Research System Card

OpenAI's February 2025 system card for deep research: an o3 variant trained with RL on browsing tasks, and its prompt-injection and exfiltration mitigations.

Contents

Summary

Explains how ChatGPT’s deep research was trained and what safety risks browsing added. Read for pages 1-4 (introduction, training, prompt injection).

Key claims

  • Deep research is powered by an early version of o3 optimised for web browsing; it searches, reads text, images and PDFs, pivots on what it finds, and can run Python on data.
  • Trained with reinforcement learning on new browsing datasets: searching, clicking, scrolling, reading files, using a sandboxed Python tool, and synthesising many sites.
  • Training tasks range from auto-gradable ones with ground-truth answers to open-ended ones graded against rubrics by a chain-of-thought grader model.
  • ChatGPT uses a second o3-mini model to summarise chains of thought.
  • Prompt injection in browsed pages can cause two harms: wrong answers (for example recommending the wrong product) and data exfiltration (for example leaking an API key from context through a network request).
  • Mitigations: safety training data against injections, plus a system-level rule that the agent cannot navigate to or construct arbitrary URLs.

Feeds

Deep research agents

Suggest changes →